Alibaba just released Qwen 3.8 27B, and it’s scoring 52 on the Artificial Analysis Intelligence Index. That’s the same score as OpenAI’s GPT-5.6 Luna (max), and just one point behind GLM-5.2 (max) and DeepSeek V4 Pro.

Here’s what makes that number ridiculous: GLM-5.2 is 753 billion parameters. DeepSeek V4 Pro is 1.7 trillion. Luna’s size is unknown, but it’s almost certainly massive. Qwen 3.8 27B is doing the same work with 27 billion parameters.

The efficiency gains here are absurd. If you’re running local inference or trying to keep cloud costs reasonable, this is the kind of model that changes your architecture decisions. You can fit this on consumer hardware. You can run it fast. And apparently, you’re not giving up much quality to do it.

This continues Alibaba’s pattern of shipping genuinely competitive open models. The Qwen series has been solid for months, but this release is different. Matching frontier commercial models at this parameter count isn’t supposed to happen yet.

Who should care: Anyone running inference at scale, anyone building on-device AI, anyone who thought “good enough” meant choosing between quality and cost. Who can ignore it: If you’re already locked into OpenAI or Anthropic’s APIs and don’t care about cost, this won’t change your life.

OpenAI’s Vision Model Actually Works Now

Speaking of OpenAI, they released GPT-5.6 Sol, and it’s the best vision model they’ve ever shipped. According to analysis from Roboflow, Sol significantly outperforms previous GPT models on vision tasks.

The details matter here. Previous GPT vision models were fine for simple tasks but unreliable for anything requiring precision. Sol apparently handles complex visual reasoning, OCR, diagram interpretation, and spatial understanding much better than GPT-4 Vision or GPT-5.6’s other variants.

This is useful if you’re building anything that needs to understand images or documents. The previous generation was good enough for demos but sketchy in production. Sol seems like it crosses that line.

Pricing and API details weren’t in the announcement, so check OpenAI’s documentation for current rates. If you tried vision models a year ago and gave up, this is worth a second look.

Anthropic Adds $18B in Two Months

Anthropic’s annualized revenue hit $65 billion, up $18 billion in two months. That’s not a typo. They added more revenue in eight weeks than most AI companies will see in their entire existence.

For context, Anthropic was at $47 billion in annualized revenue in June. This kind of growth is unusual even by AI standards. It suggests either massive enterprise adoption, significant pricing power, or both.

This matters because revenue trajectory determines who survives the next consolidation phase. Anthropic is clearly in the survivor column. If you’re building on Claude, this is good news for long-term API stability. If you’re competing with Anthropic, this is terrifying.

The Rest

Wispr raised $280 million at a $2 billion valuation and is expanding beyond dictation into meeting notes. It’s another swing at the “replace Zoom transcription” market, which is getting crowded.

Nvidia invested $1.5 billion in SoftBank’s data center developer, specifically to guarantee its chips power an OpenAI data center. The infrastructure arms race continues.

AI automation startup Relay shut down, with the team joining Google’s Chrome team. Founder Jacob Bank says they have “ambitious plans to help you work with AI in Chrome,” which probably means more Gemini integration is coming.

And in a story that’s equal parts fascinating and depressing, Amazon is reportedly destroying rare books to train AI models. These are valuable training sources precisely because they’re not already online. If you care about book preservation, this one will make you angry.